DAVID LEVIN Near-bestapproximationstothesolutionofFredholm ... · R A I R O Analyse numérique/Numerical Analysis (vol 16,11° 2, 1982, p 129 à 141) IMEAR-BEST APPROXIMATIONS TO

RAIROANALYSE NUMÉRIQUE

DAVID LEVINNear-best approximations to the solution of Fredholmintegral equation of the second kindRAIRO – Analyse numérique, tome 16, no 2 (1982), p. 129-141.<http://www.numdam.org/item?id=M2AN_1982__16_2_129_0>

© AFCET, 1982, tous droits réservés.

L’accès aux archives de la revue « RAIRO – Analyse numérique » impliquel’accord avec les conditions générales d’utilisation (http://www.numdam.org/legal.php). Toute utilisation commerciale ou impression systématique estconstitutive d’une infraction pénale. Toute copie ou impression de ce fi-chier doit contenir la présente mention de copyright.

Article numérisé dans le cadre du programmeNumérisation de documents anciens mathématiques

http://www.numdam.org/

R A I R O Analyse numérique/Numerical Analysis(vol 16,11° 2, 1982, p 129 à 141)

IMEAR-BEST APPROXIMATIONS TO THE SOLUTIONOF FREDHOLM INTEGRAL EQUATION

OF THE SECOND KIND (*)

by David LEVIN (X)

Communicated by J DOUGLAS

Résume — On considère des approximations par colîocation de la solution d'équation intégralede Fredholm de seconde espèce, et une approximation ponctuelle presque optimale est définie parcolîocation en des abscisses de formules de quadrature optimale On obtient une approximationglobale presque optimale en ajoutant des termes correcteurs à l'approximation par colîocation,a l'aide de propriétés de base du noyau resolvant Un procédé semblable aux itérations de Neumannaméliore l'approximation par colîocation même dans les cas où la série de Neumann diverge Ondiscute en détail le cas de noyaux à singularité algébrique et on donne un exemple numérique

Abstract — Colîocation approximations to the solution of Fredholm intégral équation of thesecond kind are discussed, and a pointwise near-best approximation is dejined by colîocation at theabscissae of some best quadtatur e formula A global neat-best approximation is obtained by addingsome correction term to the colîocation approximation, utihzing basic properties of the résolventkernel A procedure similar to Neumann itérations is shown to improve the colîocation approximationeven in cases when the Neumann séries diverges The case of kernels with algebraic singulanty isdiscussed in detail and a numencal example is given

1. POINTWISE NEAR-BEST APPROXIMATIONS

Consider Fredhom intégral équation of the second kind

u(x) - X K(x, t) u(t) dt = ƒ(*), ƒ e FJa

or, in operator form

u - XKu = f

where F is a given continuity class of fonctions on [a, b\

(*) Received m November 1980(x) School of Mathematical Sciences, Tel-Aviv University, Ramat-Aviv, Israël

R A I R O Analyse numénque/Numerical Analysis, 0399-0516/1982/129/$ 5 00

© Bordas-Dunod

130 D. LEVIN

For any X which is not an eigenvalue of K équation (1.1) has a unique solu-tion, which can be represented in terms of the résolvent kernel F of (1.1) as

u{x) = I T(x,t;X)f(t)dt ^(1.2)

or, in operator form

u = TJ

For the ideal case when the résolvent kernel is known we can obtain approxima-

tions to u(x) by using quadrature approximations to F(x. t ; X) f(t) dt.

In this context, for a given x e [ay b\ we introducé the class Hx of functionson [a, b\

Hx = {h\ h(t) = r(x, t; X) cj)(O, 4> e F } , (1.3)

and define the following best «-point approximation to u(x).

n

DÉFINITION 1.1: Let £ wf h(tf) be the best quadrature formula, in some sense,

r*for approximating h{t) dt for h e Hx. The best n-point approximation, in

Jathe same sense, io u(x) is defined as

«*(*)= £ wtnx,tt;X)f(tt). (1.4)

Let ej be the error functional of the above best quadrature formula, i.e.

ei(h)= [ h(t)dt- t *?Kt?). 0.5)

Then the error in the best «-point approximation can be expressed in terms ofejas

u(x) - u*(x) =e*(h) (1.6)

where h e Hx,

h(t)=r(x,t;X)f(t) te[a,b]. (1.7)

Usually the résolvent kernel F is not known and the above best «-pointapproximations cannot be used However, some near-best «-point approxima-

R.A.I.R.O. Analyse numérique/Numerical Analysis

FREDHOHM INTEGRAL EQUATION 131

tions can be obtained by mère knowledge of the abscissae of a best n-pointquadrature formula on Hx as foliows :

DÉFINITION 1 .2 : Let wl5 w2,..., un be n independentfunctions on [a, b] such that

ZEfjeF, (1.8)

and let tf, t%,..., t* be the abscissae of the best n-point quadrature formula, insome sensé, on Hx. Then the near-best n-point approximation to the solutionu{x) of the problem (1.1) at a given x e [a, b] is defined as

t *JUJ(X) (1-9)

where the OLJS are chosen such that ïï collocates the intégral équation at

* 1 > * • ! • > • • • > ln •

The collocation property of ü means that

û(tf) - X f K{tf9 t) û(t) dt = f(tf) , i = 1, 2,..., n, (1.10)Ja

or, using(1.8)and(1.9)

t a, ƒ/**)=ƒ(*•), i = l,2,..,«. (1.11)

Solubility of the System (1.11) can, in gênerai, be achieved by a proper choiceof the basis functions upj = 1, 2, „., n.

The following theorem exhibits the " near-best property " of the near-best«-point approximation, i.e. that the error in it is given by the error functionale* of the best n-point quadrature formula on Hx, operating on a function fromHx. The term "near-best" is being used in connection with the above bestn-point approximation w*(x), which also satisfies the same property (eq.(1.6)).

THEOREM 1 . 1 : Let ü(x) be a near-best n-point approximation given by défi-nition 2 .1. Then

u(x) - û(x) = e*(h - f oc, h\ (1.12)

where h and the hjs are in Hx, h given by (1.7) and

hJ(t) = r(x,t;X)fJ(t) te[a,b], j=\,2,...,n. (1.13)

vol 16, n» 2, 1982

132 D. LEVTN

Proof : The w/s can be represented by means of the résolvent kernel as :

Uj(x) = f r(x, t ; X) / / O A , 7 = 1, 2,..., /i. (1.14)

JaUsing (1.2), (1.9) and (1.14) we obtain

u(x) - ü(x) = f T(x, t; X) ïf{t) - £ a, ƒ//)

Now applying the best «-point quadrature formula on Hx (whose abscissaeare used in the définition of u) to the intégral in (1.15) we get

u{x) - û{x) = t < n x , t* ; k) [ƒ(/,*) - f a , ƒ / / , * ) ! +« = i |_ j = i J

The summation term in (1.16) vanishes by the collocation property (1.11)of û, and the proof is completed by using (1.7) and(l. 13).

In many practical applications ît appears that the kernel K(x, t) in (1.1)possesses some singularity, usually on some line in [a, b] x [a, b\ In suchcases, as it is shown in section 3, the résolvent kernel T(x9 t ; X) is also smgularon the same line.

Moreover, even for a smooth kernel, the résolvent F(JC, t ; X) always possessesa è(x - t) singularity, where 5 is the Dirac-ô fonction. This strong x-depen-dence of the singular structure of F(x, / ; X), considered as a function of t, isclaimed to be the reason for the relative failure of the collocation method toproduce efficient global approximations to the solution of Fredholm intégraléquations of the second kind. In our case the conséquence is that the near-best w-point collocation approximation û can be highly efficient only at thatpoint x to which it is assigned.

n

To show that let us suppose that a near-best approximation, w = £ &, «,

obtained for some x e [a, b], is being used to approximate u(y) at some y ^ x,Then, as in theorem 1,1 it can be shown that

u{y) - û{y) = é*(h - £ a, h\ . (1.17)

R A I R O Analyse numenque/Numencal Analysis

FREDHOHM INTEGRAL EQUATION 133

But hère h and the hjs are in the class Hr Thus, the error in û(y) is a resuit ofthe error functional of a best quadrature formula on Hx operating on a func-tion in Hy This error is not expected to be very small since best quadratureapproximations on Hx are not at all suitable for functions in Hyi whose singularstructure is completely different from that of the functions in Hx.

In the following section we show that a global near-best approximation to u,with a global near-best property of the kind (1.12), can be obtained by addingan appropriate correction term to M.

2. GLOBAL NEAR-BEST APPROXIMATIONS

Let us assume that the résolvent kernel F can be decomposed as

r(jc, * ;* . )= S(JC, *; X) + R(x, t\X) (2.1)

where S contains the "main" singularities of F on [a, b\ as specified in défini-tion 2.1 below. Let us also define some sets of functions on [a, b] associatedwith this décomposition,

ƒƒ* = { h | h{t) = R(x9 t; X) <K0, <t> e F, t e [a, b]}, x e [a, b]. (2.2)

DÉFINITION 2 . 1 : Let (2.1) be a décomposition of T such that the best n-pointquadrature formulae on the associated sets Hx are independent ofxfor x G [a, b]

n

and let £ w* h(tf) be such a best formula with abscissae

tfe[a,b], i = 1,2,...,/!.

A global nth order approximation ûc to u on [a, b] is defined as

ûc(x) = û(x) + f S(x, t ; X) [ƒ (0 - f OLJ fj(t)\ dt (2.3)

where U is the collocation approximation to u collocaling (1.1) at

tf9 i = 1,2,..., n ,

le,,

û(x)= t OLjU^x) (2.4)

vol. 16, n°2, 1982

134 D. LEVIN

where the ajs are determined by

t ex, ƒ / / * ) = ƒ ( C ) . '"= l , 2 , . . . , / i , (2 .5)

where f j = u3 — XKuf

Remark : The c in iïc stands for the correction of ü with the correction term

f S(x, t ; X) I" ƒ(0 - £ a, / / 0 l * • (2 - 6)

The évaluation of this correction term involves merely direct intégration sincethe a ƒ s appearing in it are the same as those defining ü in (2.4).

THEOREM 2 . 1 : The error in the global near-best approximation is given by

u(x) - üc{x) = e*(h - £ h\ Vx G [o, 6] (2.7)

w/zere h G H*, h3 e ƒƒ*, 7 = 1, 2,..., n ,

= R(x9t;X)f(t)9 (2.8)

(2.9)

e* zs ?/ze error functional of the associated best n-point quadrature formulaon the sets / /*, x e [a, è].

Proof : Using (2.1) in (1.15) and rearranging the terms using (2.3) weobtain

u{x) - üc{x) = f Rfc t ; X) I"ƒ(0 - £ a, / / o l dt. (2.10)

Applying the best «-point quadrature formula from définition 2.1 to theintégral in (2.10) we get

u(x) - ïïc(x) = t < *(*• '.*; *-) [/(',*) " E «i /A*.=1 L ->=1

The summation term vanishes using (2.5) and the proof is completed by using(2.8) and (2.9).

R A I R O Analyse numénque/Numencal Analysis

FREDHOHM INTEGRAL EQUATION 135

Since it is assumed in définition 2.1 that the best «-point quadrature formulathere is independent of x for x e [a, b], so are also the ajs in û and ûc, Thisallows the définition of ïtc as a global approximation. However, more impor-tant is the global nature of the near-best property (2.7) of wc, i.e., for anyx e [a, b] the error in ûc is expressed by the error functional of the best quadra-ture formula on H£ operating on a function for which this best quadratureapproximation is suited.

In the following section we present some basic properties of the résolventkernel which are relevant to the application of the pointwise and globalnear-best approximations presented above. We also outline a procedure forcalculating appropriate singular parts 5 of F for the case of kernels withalgebraic singularity.

3. THE SINGULARITIES OF THE RESOLVENT KERNEL

In operator form we have the relations u — XKu = ƒ and u = F/. Hence

Tf=u

= ƒ + XKu

= ƒ + XK(f + XKu)

= ƒ + XKf + X2 K2 ƒ + - 4- XmKmf+ Xm+1 Km^1 u,

Replacing u by Tf in the last term of the above expression we obtain an ope-rator identity for F

F = ƒ + XK + X2 K2 + - + XmKm + Jim + 1 * : m + 1 F . (3.1)

In (3.1) the product of ttyo intégral operators A and B, with kernels A(x9 i)and B(x, t) respectivelwis the intégral operator AB with the kernel

C(x,0 = | A(x9y)B(y9t)dy, (3.2)

and ƒ is the identity operator.Intégral operators are smoothing operators, and thus, in gênerai, Kj+1(x, i)

is smoother than Kj(x, t), Therefore, the most severe singularities of F(x, t ; X)are imbedded in the first terms of the express ion (3.1). The identity operatorcontributes a b(x — i) singularity to F(JC, t ; X) where 5 is the Dirac-5 function.

vol. 16, n° 2, 1982

136 D LEVIN

As an example we consider the important class of kernels K(x, i) with jumpdiscontinuities on x = t,

K(x, t) = K,(x, t)+ t fcj" ( * ° + , (3 3)

where Kx e C°°([fl5 è] x j#, b]), and fcj1* is the jump in theyth denvative of K,ie,

*£(,+, 0-^(r,d (3 4)ÖJCJ ÔX3

The use of the additional index 1 is to be clanfied belowUsing the equahty

IJa

"(x - yf+ (y -k i / '

it can be proved, by induction on n, that the kernel of the operator K " canbe represented as

K"(x, t) = Kn(x, t) + £ fef(X"OI; « > 1 , (3 6)

j = n - 1 J

where Kne Cœ([a, b] x [a, b\) The /c)"J's can be obtained recursiveiy fromthe fcj1)ss by the relations

/cjm) = 0_i j < m - 1]

fc<»0 = J ^ fcJD k^-}\ y > m - 1 f ( 3 7 )

i = 0 J

From the expression (3 1) for F and the relations in (3 6) it is clear thatF(x, t,X) is ïnfmitely smooth on [a, b] x [a, b] apart from a h(x — i) singu-lanty and jump discontinuities across x — t

The near-best approximations û (définition 1 2) are defined by means ofbest quadrature formulae on the set Hx in (1 3) Let F — Cœ[a, b], then Hx

is the class of înfinitely smooth fonctions on [a, b] apart from a ô(x — t) sin-gulanty and jump discontinuities at t = x A best «-point quadrature for-mula on this Hx can be composed of two Gaussian quadrature formulae,one on [a, x] and the other on [x, b\ and an additional term, <j)(x) (whereh(tt) = T(x, t, X,) (j>(0 is the mtegrand), to take care of the ô(x — t) singulantyo^ F Let ê£ d] dénote the error functional of the /c-pomt Gaussian quadratureformula on [c, d] Then, for a given x e [a, b] we choose kx such that the kx-

R A I R O Analyse numenque/Numencal Analysis

FREDHOHM INTEGRAL EQUATION 137

point Gaussian quadrature approximation on [a, x] is of the same order asthe (n — kx — l)-point Gaussian quadrature approximation on [x, b\ i.e.

0{#?\g)) « O(ei-1(0)) » 9 e C°"[a, b].

The near-best «-point approximation is then defined by collocation at thekx Gaussian points in [a, x\ at x, and at the n — kx — 1 Gaussian points in [x, b\From theorem 1.1 then follows :

COROLLARY 3 . 1 : The error in the near-best n-point approximation

n

u{x) = X ctjUjix)J = l

to the solution u(x) of the intégral équation (1.1) with a kernel of the form (3.3)and ƒ e C°°[as è] is given by

u(x) - û{x) = é£? (h - £ ajh) + e^l^ (h - £ oijh^ (3.8)

where h and the hjs are as in (1.12), hence when restricted to [a, x) they arein C°°[a, x) and restricted to (x, b] they are in COT(x, b\

For a global near-best approximation we need to find a suitable singularpart S of F as described in définition 2.1. A possible S can be induced from (3.1)as the kernel of the operator

Sm = I + XK + X2K2 + •- + XmKm (3.9)

where m can be chosen so that the " remainder "

Rm - Vn+1 Km+lT (3.10)

has a sufficiently smooth kernel. For kernels of the form (3.3) it can be shown,using (3.6), that the kernel of Rm has at least m — 1 continuo us derivatives.Furthermore, for this type of kernels a more practical S can be obtained byreplacing each Kn in (3.9) by its représentation (3.6), retaining only theterms which contribute to the jump discontinuities in djS/dxj,j~0, 1,..., m— 1.This pro vides us a singular part Sm of F,

m— 1 F m ~\ (x fV

j=o L^ i J j ! (3.11)

whose associated " remainder " Rm = F - Sm, just like Rm, has a kernel

vol. 16, n° 2, 1982

138 D. LEVIN

in C^1"1^, b\ for any x e [a, b], The sets H*m and H*m associated with Sm

and S^ are therefore in C(m" 1][a, b] for any x e [o, b] if F = C°°[a, è]. ChoosmgS = ^2»+i w e ê e t Hx Œ C{2n)[a, b] and obviously the /7-point Gaussianquadrature formula can be taken as a best «-point formula on Hf for anyx e [a, b]. The global nih order near-best approximation for this case is thusdefined as

üc(x) = £ ttj «/*) + f S2n+1(x, t; X) \f(t) - t ot, / / o ] A (3.12)

where the a/s are determined by (2.5) with the ff's taken as the n Gaussianpoints in [a, b\

Remark : The choice of Gaussian points as collocation points is alsosuggested in other works; see e.g. Pruess [4], However, the motivation there

n

is based on the attempt of making ƒ — YJ ajfj nea rly orthogonal to all

polynomials of degree < n. In the present work, the discussion leading tocollocation at Gaussian points emphasizes the significance of the singularstructure of K in this context, and reveals the necessity of adding a correctionterm to the collocation approximation. With this correction term we have,using theorem 2.1, the following promising result :

COROLLARY 3 . 2 : The error in the global nth order near-best approximationüc to the solution u of (1.1) with a kernel of the farm (3.3) and F = C°°[a, b]is given by

u(x) - üe(x) = e^lh - £ a, h\ Vx e [a, b] (3.13)

where h, hu h2,.-, hn e C(2n)[a, b] and e[„M is the error functional of the n-pointGaussian quadrature formula on [a, b\

The main contribution of the correction term in (3.12) is

<*j f M) (3-14)f 8(x - 0 f ƒ (0 - tt «. ƒ/')] dt = ƒ(*) -

due to the first term in S2tt+1. This correction seems to be essential even forsmooth iC's. By this simple correction we get the approximation

R A I R O Analyse numénque/Numencal Analysis

FREDHOHM INTEGRAL EQUATION 139

which can be interpreted as a one stage Neumann itération from the "point"

It is mentioned in Baker [1] that this simple correction frequently improvesexpansion approximations.

If in (3.12) the S2n +1 is replacée by the kernel of Sm of (3.9) then ûc becomessimply the mth Neumann iterate starting from û. It is important to noticethe différence between the present motivation of obtaining this approximationand the simpler motivation based on Neumann itérations. Since the presentmotivation is based solely upon the fact that the kernel Kj+ x(x, t) is smootherthan the kernel Kj(x, t\ it does not require the convergence of the Neumannseries.

It can be shown that corollary 3.2 holds for the (2n + l)th Neumanniterate from w. Therefore, even for a divergent Neumann series we expectthe first 2n + 1 itérations to improve the collocation approximation û,although further itérations might destroy this improvement

Repeated Neumann itérations are not commonly used since their compu-tation is expensive. However, for kernels of the form (3.3), the approxima-tion ïïc, using the computationally simple kernel S2n+1) has been shown toplay the same rôle as the (2 n + l)th Neumann iterate. The technique usedfor kernels of the form (3.3) can be easily extended to deal with more gêneraikernels of the form

K(x, t) = CU 0 + £ cfx - ifi (3.16)

with real exponents — 1 < r0 < rx < r2 < ... where Ce Cco([a, b] x[a, b]).Hère, a generalization of equality (3.5) should be used,

IJa

(x - y ) \ (y - t)% dy = B(r + l,s + 1) (x - i)r+

+s+1 (3 .17)

where B is the Bèta function.For other classes of kernels it might be more difficult to find a convenient

expression for the singular part S of F. However, the results obtained withthe near-best approximations for kernels of the form (3.3) indicate that thestudy of other types of singular kernels deserves a strong considération.

vol. 16, n° 2, 1982

140 D. LEVIN

4. NUMERICAL EXAMPLE

We consider the Fredholm équation

u(x)-X f K(x,t)u(t)dt=f(x) (4.1)Jo

with ƒ G C°°[0, 1] and with kernel

x(\ - t) 0 ^ x <c t

This kernel can be rewritten as

K(x9 t) = x{\ - t)-(x- t)+ , (4.3)

i.e., in the form (3.3) with Kt(x, t) = x(l - t\ k{ol) = 0, k[l) = - 1 and

kl1} = 0 for i > 2. Hence we can use (3.7) and (3.11) to compute the singularparts Sm of F.

To demonstrate the power of the corrected approximation ïïc we considercorrections to a iow order coïlocation approximation, a 4th order in thiscase. The 4-point collocation approximation ü is taken as the third degreepolynomial collocating the intégral équation at the four Gaussian pointsin [0, 1]. In ïïc, defined by (3.12), the computation of the f/s is performedanalytically by (1.8) and the correction term is approximated by using Simp-son's rule with h = 0.01.

To investigate the influence of particular singularities of F upon the cor-rection term we compute a séquence of approximations, ï£m), m = 0, 1, 2,...,corresponding to correction terms with Sm. Thus ïÜfl) takes care of the jumpdiscontinuities in the first m — 1 derivatives of F. We note that in this caseSii+i ^ S2i and therefore wc

(2i + 1) = wc(2°.

We tested the problem (4.1) with X = 1 and f(x) = x having the solution

u(x) = sin x/sin 1 .

In table 4.1 we give results at x = 0, 0.2, 0.4, 0.6, 0.8, and 1.0 computed fromthe collocation approximation ü and the corrected approximations üjo), ü{

c2)

and z?c4). These are compared with values computed from the analytic solu-

tion u(x).

R.A.I.R.O. Analyse numérique/Numerical Analysis

FREDHOHM INTEGRAL EQUATION 141

TABLE 4 .1 .

x 0.0 0.2 0.4 0.6 0.8 1.0

û(x) ~ 0.0002716940.236219378 0.462707145 0.670937092 0.852654706 0.999605471u^0)(x) 0.000000000 0.236098504 0.462781577 0.671017224 0.85250376M.000000000w<2)(x) 0.000000000 0.236097644 0.462782799 0.671018304 0.852502461 1.000000000u{*\x) 0.000000000 0.236097651 0.462782835 0.671018328 0.852502439 0.999999967u(x) 0.000000000 0.237097660 0.462782852 0.671018352 0.852502467 1.000000000

A clear improvement is already achieved by ûJ0) using only the fîrst cor-rection term. From three correct signifîcant figures in ûthe accuracy is improv-ing to five correct figures in û$°\ and to seven correct figures in ï^4).

We note that the development of the near-best approximations presentedhère is based upon the représentation (1.2) of the solution of problem (1.1).In fact, similar near-best approximations can be obtained for many problemswhose solution has a représentation of the form (1.2) (see [2]). In [3] this isdone for the harmonie Dirichlet problem.

REFERENCES

[1] C. T. H. BAKER, Numerical solution of intégral équations, L. M. DELVES, J. WALSH

eds., Clarendon Press, Oxford, 1974.[2] D. LEVIN, Near-best approximations to some problems in applied mathematics,

Report TR/66, Department of Mathematics, Brunel University, 1976.[3] D. LEVIN, Corrected collocation approximations for the harmonie Dirichlet problem,

J. Inst. Maths Applics, 26, 1980, 65-75.[4] S. A. PRUESS, Estimating the eigenvalues of Sturm-Liouvilleproblems by approximat-

ing the differential équation, SIAM J. Numer. Anal., 10, 1973, 55-68.

vol. 16, n°2, 1982